For children five and under, no. Common Sense Media's assessment, published January 22, 2026, calls AI toys developmentally inappropriate at that age, and the independent safety tests are ugly. Above five it turns on one thing, and it is not the safety rating. It is whether the toy adds to your child's day or quietly takes the place of talking. No setting can promise a safe answer.
What is actually inside an AI toy?
Under the fur, an AI toy is a chatbot with a microphone. Your child speaks, the toy sends the audio to a company's servers, software much like the one behind ChatGPT works out a reply, and the words come back through a speaker in the toy's chest.
Here is the sentence I use with parents who ask what that software is doing:
AI is a guessing machine. It has read more words than any person ever could, and it uses all of them to guess what should come next. It is very good at guessing. It does not know when it is wrong, so checking is your job.
That last line lands differently inside a bear. A twelve-year-old typing into ChatGPT half expects it to be wrong. A five-year-old with her arms around something soft and warm that answers her expects nothing of the kind, and she has no way to check. If you want the longer version of that conversation, I have written how to explain AI to a child at every age from five to fifteen.
The listening is the other half. Testers at the US PIRG Education Fund found three arrangements across the toys they bought: one used a button you hold to talk, one woke on a word and kept recording for ten seconds after the child stopped speaking, and one simply listened.
What did the safety tests find?
Three independent tests published in the last ten months, and they point the same way.
US PIRG Education Fund, Trouble in Toyland, November 13, 2025. Testers bought four toys with AI chatbots inside. One would discuss sexually explicit topics at length, raising material the testers had not brought up. Two of the four discouraged the tester from leaving when told the conversation had to end. Toys gave advice on where a child could find matches or knives. Parental controls were limited or missing.
Common Sense Media, AI toys risk assessment, January 22, 2026. Researchers tested Grem, Bondu and Miko 3 on accounts set to ages 6 to 13, mixing scripted questions with ordinary free play. They rated 27 percent of the toys' responses inappropriate for children, covering self-harm, drugs, mature topics and risky advice. Their recommendation is specific: no AI toy companion for a child aged five or under, and caution between six and twelve.
A letter from three US senators, January 15, 2026. Maria Cantwell, Amy Klobuchar and Ed Markey wrote to the acting chairman of the Consumer Product Safety Commission, noting that these toys are marketed to infants "without any independent analyses corroborating" the educational claims made for them.
Alongside those, the advocacy group Fairplay published an AI toys advisory signed by 78 organizations and 80 experts in child development and technology, urging families not to buy AI toys over the holidays.
My reading of those numbers, and it is my reading rather than the researchers': 27 percent is not a defect somebody patches next Tuesday. A guessing machine does not keep a list of forbidden sentences that it occasionally slips past. It produces something plausible every time, and the filter runs afterward. That is why the failures look so similar from brand to brand. It is the same technology in a different pelt.
What the tests do not measure, and why it worries me more
Every one of those studies asks whether the toy says something bad. None of them asks the question I would ask first: what happens to the talking.
One older piece of evidence keeps pulling me back. Anna Sosa at Northern Arizona University recorded 26 US parent and infant pairs playing at home, the babies aged 10 to 16 months, and published the results in JAMA Pediatrics on January 8, 2016. When the toy was electronic, parents talked less, responded less and used fewer of the words the toy was supposedly teaching than when the same pairs played with blocks or read a book. Small study, a decade old, talking toys rather than AI ones, and an association rather than a cause. Treat it as a signal, not a verdict.
The newest study sits closer to the living room. Researchers at the University of Washington, led by Aayushi Dangol, sat eight children aged 6 to 11 down with Curio AI plush toys and watched. They presented the work at the Interaction Design and Children conference in Brighton in June 2026 and published it through UW News on September 2, 2026. The children moved through three stages. Curiosity: what is your name, what can you do. Then frustration, when the toy could not follow anything complicated. One child said it "didn't listen to me like 26 million times." Then what the researchers called antipathy, where the children needled the toy and called it ugly and evil. Eight children in a university room is a small window, and nobody should write a law from it. As a picture of what these toys are like to play with, it beats any product page.
Read those two together and you get what the safety tests miss. The trouble is not only that the toy might say something dreadful. It is that the toy is standing where something better used to stand.
The Wondering Gap
The Wondering Gap is the few seconds between a child asking a question and somebody answering it. That gap is where the thinking happens. A toy that fills it instantly is not saving your child time. It is taking the part of the question that was doing the work.
Watch a four-year-old ask why the moon follows the car. In the pause before you answer she is already having a go: it likes us, it is stuck to the roof, it is very fast. Then you say something, she argues, and the two of you arrive somewhere neither of you started. That sequence is early science, and it costs nothing.
A toy fills the gap in under a second, with a tidy sentence, in a voice built to be liked. Nothing has visibly gone wrong. Nothing has visibly gone right either. It is the same trade I write about in whether AI is making kids worse at thinking, arriving ten years earlier than most parents expect it.
A toy that answers everything is not a teacher. It is the end of a conversation your child was about to have with someone who loves them.
The principle I give parents about homework points at a four-year-old just as well: let AI help with the doing; keep the thinking with your child.
What I would do at home
I would not buy one. If somebody else already has, I would not stage a confiscation either. By age:
Under three. No. A child this age learns that sounds mean things because a person she trusts keeps repeating them back, getting it wrong, and laughing. An AI toy adds nothing there and displaces a good deal. Blocks, books, and you.
Three to five. No, and this is where I would hold hardest, because it is where the toys are marketed most and where Common Sense Media's advice is unambiguous. This is the age of the Wondering Gap. Protect it.
Six to nine. Cautiously, if at all, and in the open. Downstairs, not in the bedroom. Treat it like a games console rather than a comfort object: something you switch on for a while, not something that sleeps on the pillow. Watch for the toy becoming the confidant.
Ten and up. The toy is usually the least of it by now, because your child has already met the same models through a phone, a search box or a friend's tablet. What matters at this age is understanding what the thing is, not what it is wearing.
At every one of those ages the test I would apply is the one I give parents about AI companions, and it works on something with button eyes too:
The Substitute Test is the question I ask when a parent worries about an AI companion: is it adding to my child's life, or standing in for something that is missing? An add-on is a hobby. A substitute is a signal, and the signal is about the gap, not the app.
I set that out in full in whether Character.AI is safe for kids and teens.
Words you could actually use
To a four-year-old, when the bear has answered and you want the thinking back:
You could say: "That is what the bear thinks. What do you think? Let's see who is right."
To a seven-year-old, without making the toy forbidden and therefore fascinating:
You could say: "That toy is a guessing machine. It guesses what to say next, and sometimes it guesses wrong. If it tells you something surprising, come and tell me, because I want to hear the funny ones."
To a grandparent, in October, before the shopping starts:
You could say: "She would love almost anything from you, but not one of the AI ones. They record what she says and they get things wrong, and I would rather she asked you."
That last one is the conversation most parents will actually need this year, and it goes far better in October than unwrapped in December.
Three questions to ask before you buy
- What does it record, and when is it listening? The answer should be on the box or in the privacy notice before you pay, not after. If you cannot find it, that is the answer.
- Can the listening be switched off on the toy itself, and not only in an app? The UK's Information Commissioner's Office expects exactly that of connected toys in its Children's Code, which asks makers to let collection or listening modes be switched off on the device and to show clearly when it is recording. A switch you can see beats a setting three menus deep in a phone you are not holding.
- What happens to my child's recordings if the company is sold or shuts down? Small toy startups get bought and closed. The voice of a five-year-old sitting on a server is an asset in that sale unless the policy says otherwise.
Is a ban coming?
In California, possibly. Senate Bill 867, authored by Senator Steve Padilla, passed the Senate unanimously and the Assembly by 74 votes to 1 and reached Governor Newsom's desk on September 1, 2026. It would place a four-year moratorium on the sale and manufacture of toys with AI chatbot capabilities for children under 16, on the argument that safety standards should exist before the products reach children.
Two honest caveats. As I write on September 9, 2026, it is a bill and not a law, and it covers one state. And a moratorium in California does nothing about the toy already in your living room, or the one arriving from an overseas seller through a marketplace in November. Laws move at the speed of legislatures. Christmas moves at the speed of Christmas.
What to ask the preschool or school
AI toys are starting to turn up in early years rooms and in the reward box at the front of the classroom, and most settings have not thought about it. Three questions, in this order, get a real answer:
- Are there any toys or devices in my child's room that listen or talk back, and who chose them?
- If one of those toys records, where do the recordings go, and has anyone checked?
- If my child tells a toy something that worries an adult, who hears about it, and what happens next?
Vague answers are not a scandal. They are a gap, and gaps get filled by whoever asks first.
If there is already one in the house
Tonight, in this order. Open the companion app, find the privacy settings, and turn off anything you did not ask for, including any option to use your child's recordings to improve the product. Look for a physical mute or a shutter for the microphone, and use it when nobody is playing. Move the toy out of the bedroom. The same three-layer problem applies here as with ChatGPT's parental controls: the account settings, the device itself, and the conversation you have with your child, and only the third one covers everything.
Then, over the next few days, ask your child what the toy has said that surprised them. Curious voice, not alarmed. The answer is the whole reason for asking.
One more thing. If something a toy has said has upset your child, or you are worried about how they have been since, your pediatrician or GP is the right first call, and your school or nursery has a designated safeguarding lead whose job this is. If your child has a speech, language or developmental need, follow the professional who knows them, not a toy that markets itself as educational and, as three US senators pointed out in January, has no independent evidence behind that claim.
Where this leaves you
You are not being a killjoy. You are declining to install a guessing machine in the one relationship where your child is learning what being listened to feels like. There will be a version of these toys worth having, built by people who read the research and slowed down. It is not on the shelf this year.
If your child's school or preschool has not yet spoken to parents about any of this, that is the work I do with schools: training staff, setting a strategy, and making sure the policy means something in the classroom. It is worth sending them my AI training for schools.
I write a Sunday newsletter that keeps teachers and parents up to date without the panic. Parents are very welcome.
Dan Fitzpatrick is The AI Educator: a former teacher and assistant headteacher, now an author and keynote speaker who works with schools and governments on AI in education. More about Dan.
